Neutron Double-Differential Cross Sections for Spallation Reactions from an ANN Model

In the quest for cleaner, safer nuclear energy, scientists are looking toward a concept known as the accelerator-driven system. Imagine a machine that does not rely on a self-sustaining chain reaction like a traditional power plant, but instead uses a powerful beam of particles to smash into a heavy metal target. This violent collision, called spallation, shatters the heavy atoms and releases a flood of neutrons. These neutrons then drive a secondary reaction that can burn up long-lived radioactive waste and generate energy from materials that would otherwise be useless. To design such a machine, engineers need to know exactly how many neutrons are produced, how fast they are moving, and in which direction they fly when the beam hits the target. This information is crucial for safety and efficiency, yet predicting it has long been a difficult challenge because the physics of these high-speed collisions is incredibly complex.

For decades, researchers have tried to simulate these events using detailed computer models based on the laws of physics. These models attempt to track every interaction between the particles inside the nucleus, but they often struggle to match real-world measurements perfectly. Sometimes the models predict too many neutrons, sometimes too few, and the differences can be larger than the uncertainties in the experiments themselves. Because the behavior of these particles depends on so many factors—the speed of the incoming beam, the type of metal being hit, and the angle at which the neutrons escape—creating a single, perfect formula has proven elusive. The complexity is such that even the most advanced physics simulations can miss subtle patterns in the data, leaving engineers without the precise numbers they need to build the next generation of nuclear reactors.

To solve this problem, a team of researchers from the Chinese Academy of Sciences and Lanzhou University has turned to a different approach: they let the data speak for itself. Instead of trying to force the physics into a rigid set of equations, they built a flexible, data-driven model using an artificial neural network. This is a type of computer program inspired by the way the human brain learns, capable of finding hidden patterns in vast amounts of information. The researchers fed their model thousands of experimental measurements from laboratories around the world, covering a wide range of target materials and beam energies. By training the network on these real-world observations, including the natural uncertainties in the measurements, they created a tool that can predict the behavior of neutrons with remarkable accuracy.

The team discovered that by carefully selecting the right pieces of information to feed into the computer, they could capture the essential physics without needing to simulate every single collision. They found that the model worked best when it was given specific details about the angle of the neutron's flight, the energy of the incoming beam, and the size and composition of the target nucleus. Surprisingly, they also found that adding certain extra details actually made the model less accurate, while leaving out others had no effect. The final, streamlined model learned to reproduce the complex shapes of the neutron energy spectra, including the sharp peaks that occur when particles bounce off each other in specific ways. These peaks are difficult for traditional physics codes to get right, but the neural network captured them naturally because it had seen them in the experimental data.

One of the most significant tests of this new model was its ability to predict results for situations it had never seen before. The researchers tested the model against experimental data for indium targets, beam energies of 597, 1000, and 1500 MeV, and neutron angles of 90 degrees, none of which were included in the training set. In these cases, the model's predictions aligned closely with the actual experimental data, demonstrating strong generalization capabilities. However, the researchers also noted a limit to this power: when they attempted to extrapolate to a much higher beam energy of 3000 MeV, the model's predictions did not agree well with the data, particularly at low neutron energies and in the high-energy tail. To further explore the model's utility, the researchers generated detailed predictions for a copper target, a material not used in the training or testing phases. These copper results are presented as a resource for future validation, rather than as a confirmed comparison against existing measurements.

The researchers also explored how to make the model robust enough to handle the noise and imperfections inherent in real experimental data. They found that trying to smooth out the data or force the model to be overly simple actually hurt its performance, causing it to miss the important peaks and valleys in the neutron energy distribution. Instead, by training the model on many slightly different versions of the data that included the natural experimental errors, the network learned to ignore the noise while keeping the true signal. This approach allowed the model to remain flexible and accurate without becoming confused by the small fluctuations that are unavoidable in any physical measurement.

This work represents a shift in how scientists approach complex nuclear problems. Rather than relying solely on theoretical simulations that may miss subtle details, the team demonstrated that a data-driven approach can learn the underlying patterns directly from nature. The resulting model is not just a theoretical exercise; it is a practical tool that can be used to improve the design of accelerator-driven systems, which are seen as a promising path for managing nuclear waste and generating energy. By providing a high-precision map of how neutrons behave during spallation, this new model helps engineers build safer and more efficient reactors. The success of this method suggests that similar data-driven techniques could be applied to other areas of nuclear physics, potentially leading to a new era where machine learning and experimental data work together to solve some of the most challenging problems in energy science.